mm/backing-dev.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2
   3 #include <linux/blkdev.h>
   4 #include <linux/wait.h>
   5 #include <linux/rbtree.h>
   6 #include <linux/kthread.h>
   7 #include <linux/backing-dev.h>
   8 #include <linux/blk-cgroup.h>
   9 #include <linux/freezer.h>
  10 #include <linux/fs.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/mm.h>
  13 #include <linux/sched/mm.h>
  14 #include <linux/sched.h>
  15 #include <linux/module.h>
  16 #include <linux/writeback.h>
  17 #include <linux/device.h>
  18 #include <trace/events/writeback.h>
  19
  20 struct backing_dev_info noop_backing_dev_info;
  21 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  22
  23 static const char *bdi_unknown_name = "(unknown)";
  24
  25 /*
  26  * bdi_lock protects bdi_tree and updates to bdi_list. bdi_list has RCU
  27  * reader side locking.
  28  */
  29 DEFINE_SPINLOCK(bdi_lock);
  30 static u64 bdi_id_cursor;
  31 static struct rb_root bdi_tree = RB_ROOT;
  32 LIST_HEAD(bdi_list);
  33
  34 /* bdi_wq serves all asynchronous writeback tasks */
  35 struct workqueue_struct *bdi_wq;
  36
  37 #define K(x) ((x) << (PAGE_SHIFT - 10))
  38
  39 #ifdef CONFIG_DEBUG_FS
  40 #include <linux/debugfs.h>
  41 #include <linux/seq_file.h>
  42
  43 static struct dentry *bdi_debug_root;
  44
  45 static void bdi_debug_init(void)
  46 {
  47         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  48 }
  49
  50 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  51 {
  52         struct backing_dev_info *bdi = m->private;
  53         struct bdi_writeback *wb = &bdi->wb;
  54         unsigned long background_thresh;
  55         unsigned long dirty_thresh;
  56         unsigned long wb_thresh;
  57         unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
  58         struct inode *inode;
  59
  60         nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
  61         spin_lock(&wb->list_lock);
  62         list_for_each_entry(inode, &wb->b_dirty, i_io_list)
  63                 nr_dirty++;
  64         list_for_each_entry(inode, &wb->b_io, i_io_list)
  65                 nr_io++;
  66         list_for_each_entry(inode, &wb->b_more_io, i_io_list)
  67                 nr_more_io++;
  68         list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
  69                 if (inode->i_state & I_DIRTY_TIME)
  70                         nr_dirty_time++;
  71         spin_unlock(&wb->list_lock);
  72
  73         global_dirty_limits(&background_thresh, &dirty_thresh);
  74         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
  75
  76         seq_printf(m,
  77                    "BdiWriteback:       %10lu kB\n"
  78                    "BdiReclaimable:     %10lu kB\n"
  79                    "BdiDirtyThresh:     %10lu kB\n"
  80                    "DirtyThresh:        %10lu kB\n"
  81                    "BackgroundThresh:   %10lu kB\n"
  82                    "BdiDirtied:         %10lu kB\n"
  83                    "BdiWritten:         %10lu kB\n"
  84                    "BdiWriteBandwidth:  %10lu kBps\n"
  85                    "b_dirty:            %10lu\n"
  86                    "b_io:               %10lu\n"
  87                    "b_more_io:          %10lu\n"
  88                    "b_dirty_time:       %10lu\n"
  89                    "bdi_list:           %10u\n"
  90                    "state:              %10lx\n",
  91                    (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
  92                    (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
  93                    K(wb_thresh),
  94                    K(dirty_thresh),
  95                    K(background_thresh),
  96                    (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
  97                    (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
  98                    (unsigned long) K(wb->write_bandwidth),
  99                    nr_dirty,
 100                    nr_io,
 101                    nr_more_io,
 102                    nr_dirty_time,
 103                    !list_empty(&bdi->bdi_list), bdi->wb.state);
 104
 105         return 0;
 106 }
 107 DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats);
 108
 109 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 110 {
 111         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 112
 113         debugfs_create_file("stats", 0444, bdi->debug_dir, bdi,
 114                             &bdi_debug_stats_fops);
 115 }
 116
 117 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 118 {
 119         debugfs_remove_recursive(bdi->debug_dir);
 120 }
 121 #else
 122 static inline void bdi_debug_init(void)
 123 {
 124 }
 125 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 126                                       const char *name)
 127 {
 128 }
 129 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 130 {
 131 }
 132 #endif
 133
 134 static ssize_t read_ahead_kb_store(struct device *dev,
 135                                   struct device_attribute *attr,
 136                                   const char *buf, size_t count)
 137 {
 138         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 139         unsigned long read_ahead_kb;
 140         ssize_t ret;
 141
 142         ret = kstrtoul(buf, 10, &read_ahead_kb);
 143         if (ret < 0)
 144                 return ret;
 145
 146         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 147
 148         return count;
 149 }
 150
 151 #define BDI_SHOW(name, expr)                                            \
 152 static ssize_t name##_show(struct device *dev,                          \
 153                            struct device_attribute *attr, char *buf)    \
 154 {                                                                       \
 155         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 156                                                                         \
 157         return sysfs_emit(buf, "%lld\n", (long long)expr);              \
 158 }                                                                       \
 159 static DEVICE_ATTR_RW(name);
 160
 161 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 162
 163 static ssize_t min_ratio_store(struct device *dev,
 164                 struct device_attribute *attr, const char *buf, size_t count)
 165 {
 166         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 167         unsigned int ratio;
 168         ssize_t ret;
 169
 170         ret = kstrtouint(buf, 10, &ratio);
 171         if (ret < 0)
 172                 return ret;
 173
 174         ret = bdi_set_min_ratio(bdi, ratio);
 175         if (!ret)
 176                 ret = count;
 177
 178         return ret;
 179 }
 180 BDI_SHOW(min_ratio, bdi->min_ratio / BDI_RATIO_SCALE)
 181
 182 static ssize_t min_ratio_fine_store(struct device *dev,
 183                 struct device_attribute *attr, const char *buf, size_t count)
 184 {
 185         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 186         unsigned int ratio;
 187         ssize_t ret;
 188
 189         ret = kstrtouint(buf, 10, &ratio);
 190         if (ret < 0)
 191                 return ret;
 192
 193         ret = bdi_set_min_ratio_no_scale(bdi, ratio);
 194         if (!ret)
 195                 ret = count;
 196
 197         return ret;
 198 }
 199 BDI_SHOW(min_ratio_fine, bdi->min_ratio)
 200
 201 static ssize_t max_ratio_store(struct device *dev,
 202                 struct device_attribute *attr, const char *buf, size_t count)
 203 {
 204         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 205         unsigned int ratio;
 206         ssize_t ret;
 207
 208         ret = kstrtouint(buf, 10, &ratio);
 209         if (ret < 0)
 210                 return ret;
 211
 212         ret = bdi_set_max_ratio(bdi, ratio);
 213         if (!ret)
 214                 ret = count;
 215
 216         return ret;
 217 }
 218 BDI_SHOW(max_ratio, bdi->max_ratio / BDI_RATIO_SCALE)
 219
 220 static ssize_t max_ratio_fine_store(struct device *dev,
 221                 struct device_attribute *attr, const char *buf, size_t count)
 222 {
 223         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 224         unsigned int ratio;
 225         ssize_t ret;
 226
 227         ret = kstrtouint(buf, 10, &ratio);
 228         if (ret < 0)
 229                 return ret;
 230
 231         ret = bdi_set_max_ratio_no_scale(bdi, ratio);
 232         if (!ret)
 233                 ret = count;
 234
 235         return ret;
 236 }
 237 BDI_SHOW(max_ratio_fine, bdi->max_ratio)
 238
 239 static ssize_t min_bytes_show(struct device *dev,
 240                               struct device_attribute *attr,
 241                               char *buf)
 242 {
 243         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 244
 245         return sysfs_emit(buf, "%llu\n", bdi_get_min_bytes(bdi));
 246 }
 247
 248 static ssize_t min_bytes_store(struct device *dev,
 249                 struct device_attribute *attr, const char *buf, size_t count)
 250 {
 251         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 252         u64 bytes;
 253         ssize_t ret;
 254
 255         ret = kstrtoull(buf, 10, &bytes);
 256         if (ret < 0)
 257                 return ret;
 258
 259         ret = bdi_set_min_bytes(bdi, bytes);
 260         if (!ret)
 261                 ret = count;
 262
 263         return ret;
 264 }
 265 static DEVICE_ATTR_RW(min_bytes);
 266
 267 static ssize_t max_bytes_show(struct device *dev,
 268                               struct device_attribute *attr,
 269                               char *buf)
 270 {
 271         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 272
 273         return sysfs_emit(buf, "%llu\n", bdi_get_max_bytes(bdi));
 274 }
 275
 276 static ssize_t max_bytes_store(struct device *dev,
 277                 struct device_attribute *attr, const char *buf, size_t count)
 278 {
 279         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 280         u64 bytes;
 281         ssize_t ret;
 282
 283         ret = kstrtoull(buf, 10, &bytes);
 284         if (ret < 0)
 285                 return ret;
 286
 287         ret = bdi_set_max_bytes(bdi, bytes);
 288         if (!ret)
 289                 ret = count;
 290
 291         return ret;
 292 }
 293 static DEVICE_ATTR_RW(max_bytes);
 294
 295 static ssize_t stable_pages_required_show(struct device *dev,
 296                                           struct device_attribute *attr,
 297                                           char *buf)
 298 {
 299         dev_warn_once(dev,
 300                 "the stable_pages_required attribute has been removed. Use the stable_writes queue attribute instead.\n");
 301         return sysfs_emit(buf, "%d\n", 0);
 302 }
 303 static DEVICE_ATTR_RO(stable_pages_required);
 304
 305 static ssize_t strict_limit_store(struct device *dev,
 306                 struct device_attribute *attr, const char *buf, size_t count)
 307 {
 308         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 309         unsigned int strict_limit;
 310         ssize_t ret;
 311
 312         ret = kstrtouint(buf, 10, &strict_limit);
 313         if (ret < 0)
 314                 return ret;
 315
 316         ret = bdi_set_strict_limit(bdi, strict_limit);
 317         if (!ret)
 318                 ret = count;
 319
 320         return ret;
 321 }
 322
 323 static ssize_t strict_limit_show(struct device *dev,
 324                 struct device_attribute *attr, char *buf)
 325 {
 326         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 327
 328         return sysfs_emit(buf, "%d\n",
 329                         !!(bdi->capabilities & BDI_CAP_STRICTLIMIT));
 330 }
 331 static DEVICE_ATTR_RW(strict_limit);
 332
 333 static struct attribute *bdi_dev_attrs[] = {
 334         &dev_attr_read_ahead_kb.attr,
 335         &dev_attr_min_ratio.attr,
 336         &dev_attr_min_ratio_fine.attr,
 337         &dev_attr_max_ratio.attr,
 338         &dev_attr_max_ratio_fine.attr,
 339         &dev_attr_min_bytes.attr,
 340         &dev_attr_max_bytes.attr,
 341         &dev_attr_stable_pages_required.attr,
 342         &dev_attr_strict_limit.attr,
 343         NULL,
 344 };
 345 ATTRIBUTE_GROUPS(bdi_dev);
 346
 347 static const struct class bdi_class = {
 348         .name           = "bdi",
 349         .dev_groups     = bdi_dev_groups,
 350 };
 351
 352 static __init int bdi_class_init(void)
 353 {
 354         int ret;
 355
 356         ret = class_register(&bdi_class);
 357         if (ret)
 358                 return ret;
 359
 360         bdi_debug_init();
 361
 362         return 0;
 363 }
 364 postcore_initcall(bdi_class_init);
 365
 366 static int __init default_bdi_init(void)
 367 {
 368         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
 369                                  WQ_SYSFS, 0);
 370         if (!bdi_wq)
 371                 return -ENOMEM;
 372         return 0;
 373 }
 374 subsys_initcall(default_bdi_init);
 375
 376 /*
 377  * This function is used when the first inode for this wb is marked dirty. It
 378  * wakes-up the corresponding bdi thread which should then take care of the
 379  * periodic background write-out of dirty inodes. Since the write-out would
 380  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 381  * set up a timer which wakes the bdi thread up later.
 382  *
 383  * Note, we wouldn't bother setting up the timer, but this function is on the
 384  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 385  * by delaying the wake-up.
 386  *
 387  * We have to be careful not to postpone flush work if it is scheduled for
 388  * earlier. Thus we use queue_delayed_work().
 389  */
 390 void wb_wakeup_delayed(struct bdi_writeback *wb)
 391 {
 392         unsigned long timeout;
 393
 394         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 395         spin_lock_irq(&wb->work_lock);
 396         if (test_bit(WB_registered, &wb->state))
 397                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 398         spin_unlock_irq(&wb->work_lock);
 399 }
 400
 401 static void wb_update_bandwidth_workfn(struct work_struct *work)
 402 {
 403         struct bdi_writeback *wb = container_of(to_delayed_work(work),
 404                                                 struct bdi_writeback, bw_dwork);
 405
 406         wb_update_bandwidth(wb);
 407 }
 408
 409 /*
 410  * Initial write bandwidth: 100 MB/s
 411  */
 412 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 413
 414 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 415                    gfp_t gfp)
 416 {
 417         int i, err;
 418
 419         memset(wb, 0, sizeof(*wb));
 420
 421         wb->bdi = bdi;
 422         wb->last_old_flush = jiffies;
 423         INIT_LIST_HEAD(&wb->b_dirty);
 424         INIT_LIST_HEAD(&wb->b_io);
 425         INIT_LIST_HEAD(&wb->b_more_io);
 426         INIT_LIST_HEAD(&wb->b_dirty_time);
 427         spin_lock_init(&wb->list_lock);
 428
 429         atomic_set(&wb->writeback_inodes, 0);
 430         wb->bw_time_stamp = jiffies;
 431         wb->balanced_dirty_ratelimit = INIT_BW;
 432         wb->dirty_ratelimit = INIT_BW;
 433         wb->write_bandwidth = INIT_BW;
 434         wb->avg_write_bandwidth = INIT_BW;
 435
 436         spin_lock_init(&wb->work_lock);
 437         INIT_LIST_HEAD(&wb->work_list);
 438         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 439         INIT_DELAYED_WORK(&wb->bw_dwork, wb_update_bandwidth_workfn);
 440         wb->dirty_sleep = jiffies;
 441
 442         err = fprop_local_init_percpu(&wb->completions, gfp);
 443         if (err)
 444                 return err;
 445
 446         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 447                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 448                 if (err)
 449                         goto out_destroy_stat;
 450         }
 451
 452         return 0;
 453
 454 out_destroy_stat:
 455         while (i--)
 456                 percpu_counter_destroy(&wb->stat[i]);
 457         fprop_local_destroy_percpu(&wb->completions);
 458         return err;
 459 }
 460
 461 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 462
 463 /*
 464  * Remove bdi from the global list and shutdown any threads we have running
 465  */
 466 static void wb_shutdown(struct bdi_writeback *wb)
 467 {
 468         /* Make sure nobody queues further work */
 469         spin_lock_irq(&wb->work_lock);
 470         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 471                 spin_unlock_irq(&wb->work_lock);
 472                 return;
 473         }
 474         spin_unlock_irq(&wb->work_lock);
 475
 476         cgwb_remove_from_bdi_list(wb);
 477         /*
 478          * Drain work list and shutdown the delayed_work.  !WB_registered
 479          * tells wb_workfn() that @wb is dying and its work_list needs to
 480          * be drained no matter what.
 481          */
 482         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 483         flush_delayed_work(&wb->dwork);
 484         WARN_ON(!list_empty(&wb->work_list));
 485         flush_delayed_work(&wb->bw_dwork);
 486 }
 487
 488 static void wb_exit(struct bdi_writeback *wb)
 489 {
 490         int i;
 491
 492         WARN_ON(delayed_work_pending(&wb->dwork));
 493
 494         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 495                 percpu_counter_destroy(&wb->stat[i]);
 496
 497         fprop_local_destroy_percpu(&wb->completions);
 498 }
 499
 500 #ifdef CONFIG_CGROUP_WRITEBACK
 501
 502 #include <linux/memcontrol.h>
 503
 504 /*
 505  * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, offline_cgwbs and
 506  * memcg->cgwb_list.  bdi->cgwb_tree is also RCU protected.
 507  */
 508 static DEFINE_SPINLOCK(cgwb_lock);
 509 static struct workqueue_struct *cgwb_release_wq;
 510
 511 static LIST_HEAD(offline_cgwbs);
 512 static void cleanup_offline_cgwbs_workfn(struct work_struct *work);
 513 static DECLARE_WORK(cleanup_offline_cgwbs_work, cleanup_offline_cgwbs_workfn);
 514
 515 static void cgwb_free_rcu(struct rcu_head *rcu_head)
 516 {
 517         struct bdi_writeback *wb = container_of(rcu_head,
 518                         struct bdi_writeback, rcu);
 519
 520         percpu_ref_exit(&wb->refcnt);
 521         kfree(wb);
 522 }
 523
 524 static void cgwb_release_workfn(struct work_struct *work)
 525 {
 526         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 527                                                 release_work);
 528         struct backing_dev_info *bdi = wb->bdi;
 529
 530         mutex_lock(&wb->bdi->cgwb_release_mutex);
 531         wb_shutdown(wb);
 532
 533         css_put(wb->memcg_css);
 534         css_put(wb->blkcg_css);
 535         mutex_unlock(&wb->bdi->cgwb_release_mutex);
 536
 537         /* triggers blkg destruction if no online users left */
 538         blkcg_unpin_online(wb->blkcg_css);
 539
 540         fprop_local_destroy_percpu(&wb->memcg_completions);
 541
 542         spin_lock_irq(&cgwb_lock);
 543         list_del(&wb->offline_node);
 544         spin_unlock_irq(&cgwb_lock);
 545
 546         wb_exit(wb);
 547         bdi_put(bdi);
 548         WARN_ON_ONCE(!list_empty(&wb->b_attached));
 549         call_rcu(&wb->rcu, cgwb_free_rcu);
 550 }
 551
 552 static void cgwb_release(struct percpu_ref *refcnt)
 553 {
 554         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 555                                                 refcnt);
 556         queue_work(cgwb_release_wq, &wb->release_work);
 557 }
 558
 559 static void cgwb_kill(struct bdi_writeback *wb)
 560 {
 561         lockdep_assert_held(&cgwb_lock);
 562
 563         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 564         list_del(&wb->memcg_node);
 565         list_del(&wb->blkcg_node);
 566         list_add(&wb->offline_node, &offline_cgwbs);
 567         percpu_ref_kill(&wb->refcnt);
 568 }
 569
 570 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 571 {
 572         spin_lock_irq(&cgwb_lock);
 573         list_del_rcu(&wb->bdi_node);
 574         spin_unlock_irq(&cgwb_lock);
 575 }
 576
 577 static int cgwb_create(struct backing_dev_info *bdi,
 578                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 579 {
 580         struct mem_cgroup *memcg;
 581         struct cgroup_subsys_state *blkcg_css;
 582         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 583         struct bdi_writeback *wb;
 584         unsigned long flags;
 585         int ret = 0;
 586
 587         memcg = mem_cgroup_from_css(memcg_css);
 588         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 589         memcg_cgwb_list = &memcg->cgwb_list;
 590         blkcg_cgwb_list = blkcg_get_cgwb_list(blkcg_css);
 591
 592         /* look up again under lock and discard on blkcg mismatch */
 593         spin_lock_irqsave(&cgwb_lock, flags);
 594         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 595         if (wb && wb->blkcg_css != blkcg_css) {
 596                 cgwb_kill(wb);
 597                 wb = NULL;
 598         }
 599         spin_unlock_irqrestore(&cgwb_lock, flags);
 600         if (wb)
 601                 goto out_put;
 602
 603         /* need to create a new one */
 604         wb = kmalloc(sizeof(*wb), gfp);
 605         if (!wb) {
 606                 ret = -ENOMEM;
 607                 goto out_put;
 608         }
 609
 610         ret = wb_init(wb, bdi, gfp);
 611         if (ret)
 612                 goto err_free;
 613
 614         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 615         if (ret)
 616                 goto err_wb_exit;
 617
 618         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 619         if (ret)
 620                 goto err_ref_exit;
 621
 622         wb->memcg_css = memcg_css;
 623         wb->blkcg_css = blkcg_css;
 624         INIT_LIST_HEAD(&wb->b_attached);
 625         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 626         set_bit(WB_registered, &wb->state);
 627         bdi_get(bdi);
 628
 629         /*
 630          * The root wb determines the registered state of the whole bdi and
 631          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 632          * whether they're still online.  Don't link @wb if any is dead.
 633          * See wb_memcg_offline() and wb_blkcg_offline().
 634          */
 635         ret = -ENODEV;
 636         spin_lock_irqsave(&cgwb_lock, flags);
 637         if (test_bit(WB_registered, &bdi->wb.state) &&
 638             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 639                 /* we might have raced another instance of this function */
 640                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 641                 if (!ret) {
 642                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 643                         list_add(&wb->memcg_node, memcg_cgwb_list);
 644                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 645                         blkcg_pin_online(blkcg_css);
 646                         css_get(memcg_css);
 647                         css_get(blkcg_css);
 648                 }
 649         }
 650         spin_unlock_irqrestore(&cgwb_lock, flags);
 651         if (ret) {
 652                 if (ret == -EEXIST)
 653                         ret = 0;
 654                 goto err_fprop_exit;
 655         }
 656         goto out_put;
 657
 658 err_fprop_exit:
 659         bdi_put(bdi);
 660         fprop_local_destroy_percpu(&wb->memcg_completions);
 661 err_ref_exit:
 662         percpu_ref_exit(&wb->refcnt);
 663 err_wb_exit:
 664         wb_exit(wb);
 665 err_free:
 666         kfree(wb);
 667 out_put:
 668         css_put(blkcg_css);
 669         return ret;
 670 }
 671
 672 /**
 673  * wb_get_lookup - get wb for a given memcg
 674  * @bdi: target bdi
 675  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 676  *
 677  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
 678  * refcount incremented.
 679  *
 680  * This function uses css_get() on @memcg_css and thus expects its refcnt
 681  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 682  * @memcg_css isn't enough.  try_get it before calling this function.
 683  *
 684  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 685  * memcg on the default hierarchy, memcg association is guaranteed to be
 686  * more specific (equal or descendant to the associated blkcg) and thus can
 687  * identify both the memcg and blkcg associations.
 688  *
 689  * Because the blkcg associated with a memcg may change as blkcg is enabled
 690  * and disabled closer to root in the hierarchy, each wb keeps track of
 691  * both the memcg and blkcg associated with it and verifies the blkcg on
 692  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 693  * created.
 694  */
 695 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
 696                                     struct cgroup_subsys_state *memcg_css)
 697 {
 698         struct bdi_writeback *wb;
 699
 700         if (!memcg_css->parent)
 701                 return &bdi->wb;
 702
 703         rcu_read_lock();
 704         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 705         if (wb) {
 706                 struct cgroup_subsys_state *blkcg_css;
 707
 708                 /* see whether the blkcg association has changed */
 709                 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 710                 if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
 711                         wb = NULL;
 712                 css_put(blkcg_css);
 713         }
 714         rcu_read_unlock();
 715
 716         return wb;
 717 }
 718
 719 /**
 720  * wb_get_create - get wb for a given memcg, create if necessary
 721  * @bdi: target bdi
 722  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 723  * @gfp: allocation mask to use
 724  *
 725  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 726  * create one.  See wb_get_lookup() for more details.
 727  */
 728 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 729                                     struct cgroup_subsys_state *memcg_css,
 730                                     gfp_t gfp)
 731 {
 732         struct bdi_writeback *wb;
 733
 734         might_alloc(gfp);
 735
 736         if (!memcg_css->parent)
 737                 return &bdi->wb;
 738
 739         do {
 740                 wb = wb_get_lookup(bdi, memcg_css);
 741         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 742
 743         return wb;
 744 }
 745
 746 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 747 {
 748         int ret;
 749
 750         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 751         mutex_init(&bdi->cgwb_release_mutex);
 752         init_rwsem(&bdi->wb_switch_rwsem);
 753
 754         ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
 755         if (!ret) {
 756                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 757                 bdi->wb.blkcg_css = blkcg_root_css;
 758         }
 759         return ret;
 760 }
 761
 762 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 763 {
 764         struct radix_tree_iter iter;
 765         void **slot;
 766         struct bdi_writeback *wb;
 767
 768         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 769
 770         spin_lock_irq(&cgwb_lock);
 771         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 772                 cgwb_kill(*slot);
 773         spin_unlock_irq(&cgwb_lock);
 774
 775         mutex_lock(&bdi->cgwb_release_mutex);
 776         spin_lock_irq(&cgwb_lock);
 777         while (!list_empty(&bdi->wb_list)) {
 778                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 779                                       bdi_node);
 780                 spin_unlock_irq(&cgwb_lock);
 781                 wb_shutdown(wb);
 782                 spin_lock_irq(&cgwb_lock);
 783         }
 784         spin_unlock_irq(&cgwb_lock);
 785         mutex_unlock(&bdi->cgwb_release_mutex);
 786 }
 787
 788 /*
 789  * cleanup_offline_cgwbs_workfn - try to release dying cgwbs
 790  *
 791  * Try to release dying cgwbs by switching attached inodes to the nearest
 792  * living ancestor's writeback. Processed wbs are placed at the end
 793  * of the list to guarantee the forward progress.
 794  */
 795 static void cleanup_offline_cgwbs_workfn(struct work_struct *work)
 796 {
 797         struct bdi_writeback *wb;
 798         LIST_HEAD(processed);
 799
 800         spin_lock_irq(&cgwb_lock);
 801
 802         while (!list_empty(&offline_cgwbs)) {
 803                 wb = list_first_entry(&offline_cgwbs, struct bdi_writeback,
 804                                       offline_node);
 805                 list_move(&wb->offline_node, &processed);
 806
 807                 /*
 808                  * If wb is dirty, cleaning up the writeback by switching
 809                  * attached inodes will result in an effective removal of any
 810                  * bandwidth restrictions, which isn't the goal.  Instead,
 811                  * it can be postponed until the next time, when all io
 812                  * will be likely completed.  If in the meantime some inodes
 813                  * will get re-dirtied, they should be eventually switched to
 814                  * a new cgwb.
 815                  */
 816                 if (wb_has_dirty_io(wb))
 817                         continue;
 818
 819                 if (!wb_tryget(wb))
 820                         continue;
 821
 822                 spin_unlock_irq(&cgwb_lock);
 823                 while (cleanup_offline_cgwb(wb))
 824                         cond_resched();
 825                 spin_lock_irq(&cgwb_lock);
 826
 827                 wb_put(wb);
 828         }
 829
 830         if (!list_empty(&processed))
 831                 list_splice_tail(&processed, &offline_cgwbs);
 832
 833         spin_unlock_irq(&cgwb_lock);
 834 }
 835
 836 /**
 837  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 838  * @memcg: memcg being offlined
 839  *
 840  * Also prevents creation of any new wb's associated with @memcg.
 841  */
 842 void wb_memcg_offline(struct mem_cgroup *memcg)
 843 {
 844         struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
 845         struct bdi_writeback *wb, *next;
 846
 847         spin_lock_irq(&cgwb_lock);
 848         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 849                 cgwb_kill(wb);
 850         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 851         spin_unlock_irq(&cgwb_lock);
 852
 853         queue_work(system_unbound_wq, &cleanup_offline_cgwbs_work);
 854 }
 855
 856 /**
 857  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 858  * @css: blkcg being offlined
 859  *
 860  * Also prevents creation of any new wb's associated with @blkcg.
 861  */
 862 void wb_blkcg_offline(struct cgroup_subsys_state *css)
 863 {
 864         struct bdi_writeback *wb, *next;
 865         struct list_head *list = blkcg_get_cgwb_list(css);
 866
 867         spin_lock_irq(&cgwb_lock);
 868         list_for_each_entry_safe(wb, next, list, blkcg_node)
 869                 cgwb_kill(wb);
 870         list->next = NULL;      /* prevent new wb's */
 871         spin_unlock_irq(&cgwb_lock);
 872 }
 873
 874 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 875 {
 876         spin_lock_irq(&cgwb_lock);
 877         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 878         spin_unlock_irq(&cgwb_lock);
 879 }
 880
 881 static int __init cgwb_init(void)
 882 {
 883         /*
 884          * There can be many concurrent release work items overwhelming
 885          * system_wq.  Put them in a separate wq and limit concurrency.
 886          * There's no point in executing many of these in parallel.
 887          */
 888         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 889         if (!cgwb_release_wq)
 890                 return -ENOMEM;
 891
 892         return 0;
 893 }
 894 subsys_initcall(cgwb_init);
 895
 896 #else   /* CONFIG_CGROUP_WRITEBACK */
 897
 898 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 899 {
 900         return wb_init(&bdi->wb, bdi, GFP_KERNEL);
 901 }
 902
 903 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 904
 905 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 906 {
 907         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 908 }
 909
 910 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 911 {
 912         list_del_rcu(&wb->bdi_node);
 913 }
 914
 915 #endif  /* CONFIG_CGROUP_WRITEBACK */
 916
 917 int bdi_init(struct backing_dev_info *bdi)
 918 {
 919         bdi->dev = NULL;
 920
 921         kref_init(&bdi->refcnt);
 922         bdi->min_ratio = 0;
 923         bdi->max_ratio = 100 * BDI_RATIO_SCALE;
 924         bdi->max_prop_frac = FPROP_FRAC_BASE;
 925         INIT_LIST_HEAD(&bdi->bdi_list);
 926         INIT_LIST_HEAD(&bdi->wb_list);
 927         init_waitqueue_head(&bdi->wb_waitq);
 928
 929         return cgwb_bdi_init(bdi);
 930 }
 931
 932 struct backing_dev_info *bdi_alloc(int node_id)
 933 {
 934         struct backing_dev_info *bdi;
 935
 936         bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
 937         if (!bdi)
 938                 return NULL;
 939
 940         if (bdi_init(bdi)) {
 941                 kfree(bdi);
 942                 return NULL;
 943         }
 944         bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
 945         bdi->ra_pages = VM_READAHEAD_PAGES;
 946         bdi->io_pages = VM_READAHEAD_PAGES;
 947         timer_setup(&bdi->laptop_mode_wb_timer, laptop_mode_timer_fn, 0);
 948         return bdi;
 949 }
 950 EXPORT_SYMBOL(bdi_alloc);
 951
 952 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
 953 {
 954         struct rb_node **p = &bdi_tree.rb_node;
 955         struct rb_node *parent = NULL;
 956         struct backing_dev_info *bdi;
 957
 958         lockdep_assert_held(&bdi_lock);
 959
 960         while (*p) {
 961                 parent = *p;
 962                 bdi = rb_entry(parent, struct backing_dev_info, rb_node);
 963
 964                 if (bdi->id > id)
 965                         p = &(*p)->rb_left;
 966                 else if (bdi->id < id)
 967                         p = &(*p)->rb_right;
 968                 else
 969                         break;
 970         }
 971
 972         if (parentp)
 973                 *parentp = parent;
 974         return p;
 975 }
 976
 977 /**
 978  * bdi_get_by_id - lookup and get bdi from its id
 979  * @id: bdi id to lookup
 980  *
 981  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
 982  * doesn't exist or is already unregistered.
 983  */
 984 struct backing_dev_info *bdi_get_by_id(u64 id)
 985 {
 986         struct backing_dev_info *bdi = NULL;
 987         struct rb_node **p;
 988
 989         spin_lock_bh(&bdi_lock);
 990         p = bdi_lookup_rb_node(id, NULL);
 991         if (*p) {
 992                 bdi = rb_entry(*p, struct backing_dev_info, rb_node);
 993                 bdi_get(bdi);
 994         }
 995         spin_unlock_bh(&bdi_lock);
 996
 997         return bdi;
 998 }
 999
1000 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
1001 {
1002         struct device *dev;
1003         struct rb_node *parent, **p;
1004
1005         if (bdi->dev)   /* The driver needs to use separate queues per device */
1006                 return 0;
1007
1008         vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
1009         dev = device_create(&bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
1010         if (IS_ERR(dev))
1011                 return PTR_ERR(dev);
1012
1013         cgwb_bdi_register(bdi);
1014         bdi->dev = dev;
1015
1016         bdi_debug_register(bdi, dev_name(dev));
1017         set_bit(WB_registered, &bdi->wb.state);
1018
1019         spin_lock_bh(&bdi_lock);
1020
1021         bdi->id = ++bdi_id_cursor;
1022
1023         p = bdi_lookup_rb_node(bdi->id, &parent);
1024         rb_link_node(&bdi->rb_node, parent, p);
1025         rb_insert_color(&bdi->rb_node, &bdi_tree);
1026
1027         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
1028
1029         spin_unlock_bh(&bdi_lock);
1030
1031         trace_writeback_bdi_register(bdi);
1032         return 0;
1033 }
1034
1035 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
1036 {
1037         va_list args;
1038         int ret;
1039
1040         va_start(args, fmt);
1041         ret = bdi_register_va(bdi, fmt, args);
1042         va_end(args);
1043         return ret;
1044 }
1045 EXPORT_SYMBOL(bdi_register);
1046
1047 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
1048 {
1049         WARN_ON_ONCE(bdi->owner);
1050         bdi->owner = owner;
1051         get_device(owner);
1052 }
1053
1054 /*
1055  * Remove bdi from bdi_list, and ensure that it is no longer visible
1056  */
1057 static void bdi_remove_from_list(struct backing_dev_info *bdi)
1058 {
1059         spin_lock_bh(&bdi_lock);
1060         rb_erase(&bdi->rb_node, &bdi_tree);
1061         list_del_rcu(&bdi->bdi_list);
1062         spin_unlock_bh(&bdi_lock);
1063
1064         synchronize_rcu_expedited();
1065 }
1066
1067 void bdi_unregister(struct backing_dev_info *bdi)
1068 {
1069         del_timer_sync(&bdi->laptop_mode_wb_timer);
1070
1071         /* make sure nobody finds us on the bdi_list anymore */
1072         bdi_remove_from_list(bdi);
1073         wb_shutdown(&bdi->wb);
1074         cgwb_bdi_unregister(bdi);
1075
1076         /*
1077          * If this BDI's min ratio has been set, use bdi_set_min_ratio() to
1078          * update the global bdi_min_ratio.
1079          */
1080         if (bdi->min_ratio)
1081                 bdi_set_min_ratio(bdi, 0);
1082
1083         if (bdi->dev) {
1084                 bdi_debug_unregister(bdi);
1085                 device_unregister(bdi->dev);
1086                 bdi->dev = NULL;
1087         }
1088
1089         if (bdi->owner) {
1090                 put_device(bdi->owner);
1091                 bdi->owner = NULL;
1092         }
1093 }
1094 EXPORT_SYMBOL(bdi_unregister);
1095
1096 static void release_bdi(struct kref *ref)
1097 {
1098         struct backing_dev_info *bdi =
1099                         container_of(ref, struct backing_dev_info, refcnt);
1100
1101         WARN_ON_ONCE(test_bit(WB_registered, &bdi->wb.state));
1102         WARN_ON_ONCE(bdi->dev);
1103         wb_exit(&bdi->wb);
1104         kfree(bdi);
1105 }
1106
1107 void bdi_put(struct backing_dev_info *bdi)
1108 {
1109         kref_put(&bdi->refcnt, release_bdi);
1110 }
1111 EXPORT_SYMBOL(bdi_put);
1112
1113 struct backing_dev_info *inode_to_bdi(struct inode *inode)
1114 {
1115         struct super_block *sb;
1116
1117         if (!inode)
1118                 return &noop_backing_dev_info;
1119
1120         sb = inode->i_sb;
1121 #ifdef CONFIG_BLOCK
1122         if (sb_is_blkdev_sb(sb))
1123                 return I_BDEV(inode)->bd_disk->bdi;
1124 #endif
1125         return sb->s_bdi;
1126 }
1127 EXPORT_SYMBOL(inode_to_bdi);
1128
1129 const char *bdi_dev_name(struct backing_dev_info *bdi)
1130 {
1131         if (!bdi || !bdi->dev)
1132                 return bdi_unknown_name;
1133         return bdi->dev_name;
1134 }
1135 EXPORT_SYMBOL_GPL(bdi_dev_name);